AI in engineering: what's possible vs what's actually happening

The gap: what's possible vs what's deployed

The results so far come from narrow AI with outputs you can check, such as predicting a failure or spotting a defect. Generative and agentic AI can do far more in a demo, but very little of it has reached day-to-day engineering work.

McKinsey's State of AI survey from November 2025 found that only about a third of organisations have scaled AI beyond pilots. The World Economic Forum and McKinsey named the problem in 2018, writing about Industry 4.0: pilot purgatory, where a project proves a point and then stalls. The term still fits.

Most of the gap sits outside the model, in data access, in who is accountable for the output and in the review step. Generative and agentic AI will get there when organisations redesign those three things around it. That is engineering work, and this profession is well placed to do it.

What generative and agentic AI can genuinely do

It helps to split the tools into three levels. Each step up, the AI does more on its own and the checking gets harder.

AI assistant. Copilot or ChatGPT as most people use them. It drafts, summarises and answers from what you paste in. Good for first drafts, rewording and explaining. It fails on facts it can't see.

AI connected to your data. The same model, but it looks things up in your specifications, standards and records first and cites them. This is where the value for engineering starts. It fails when the data is scattered, messy or locked away.

AI agent. It plans steps, calls tools and acts on your systems within set limits. Good at routine multi-step tasks. It fails on long chains where small errors compound, on vague goals and on 3D and spatial reasoning. MIT researchers call the right setting "bounded autonomy": tight scope, human validation and no handover of engineering accountability.

What's actually deployed

Most organisations are still experimenting. S&P Global Market Intelligence found that 42% of companies abandoned most of their AI initiatives in 2025, more than double the 17% the year before.

Inside engineering teams, the pattern I see at workshops is the Copilot gap. Licences are bought, but usage sits in email, meeting summaries and summarising long documents. Calculations, drawings, specifications and reviews are mostly untouched. The managers who would champion the next step can't build a business case for premium licences on that.

The opportunity is there once you map the work. One discovery exercise with a UK manufacturer turned up 53 candidate use cases, most of them routine handling of documents and information. The power users I train already do 19 distinct things with AI across design, data, finding information and writing. The cheat sheet lists all of them.

Where the measurable results are

The evidence gets weaker as the technology gets newer. The deployments that work share three features: a narrow task, a measurable output and verification built in.

Proven: machine learning and vision. Deutsche Bahn and KONUX predict switch failures with over 90% accuracy, verified through DB's own procedure. BMW Group has used AI vision inspection in series production since 2018. ZF detected 99% of honing tool fractures at a gearbox plant, the company's own figure, published by PwC in 2020.

Emerging: generative and agentic AI. Siemens reports 25% less reactive maintenance time from its maintenance copilot, but that comes from first pilots and is the vendor's own figure. Gartner forecast in June 2025 that over 40% of agentic AI projects will be cancelled by the end of 2027.

Frontier. Buyers are starting to ask for agents built on their own documents, and prototypes exist. Published, verified results for these in engineering are hard to find so far.

The blockers: data, accountability, the review step

Data. Engineering data sits in CAD, PLM and asset systems. Schneider Electric agreed in October 2026 to buy PTC, maker of Creo and Windchill, for $22.6bn, with the deal expected to close by Q3 2027. Whether your AI can reach that data, and on whose terms, is now a commercial and governance question. In MIT's 2026 study, one NASA engineer found mission knowledge spread across 26 data sources.

Accountability. TfL's AI trial at Willesden Green ran 11 algorithms on station CCTV from October 2022 to September 2023. It raised more than 44,000 alerts and sent about 19,000 to staff in real time. How many were acted on, or right, is largely redacted. Who decides what an alert means, and who answers when it's wrong?

The review step. AI makes the first draft almost free. Checking it costs the same as before, and the person signing off is still accountable. Without a redesign you get rubber-stamping or no net time saved.

Five practices of organisations that get past pilot

These follow the life of a project, from scoping to roll-out.

1. Map the workflow first. Before choosing a tool, map who does what, where the information comes from and where it waits.
2. Pick checkable tasks. Start where the output can be verified and mistakes are cheap.
3. Design the review step. Decide who checks, against what, and who signs off. Put it inside the design reviews you already run.
4. Measure after review. Count the time saved once checking is included. A business case built on drafting time alone won't hold up.
5. Protect judgement. Engineers learn judgement from a senior's red pen. If AI writes the first draft and juniors only approve it, that loop can disappear. Ask juniors to explain an AI output before accepting it, use AI drafts as review practice and keep some work unassisted.

Practices 2 and 3 match what MIT researchers found in interviews with 28 engineering and manufacturing firms in 2026.

Questions people ask

About the speaker

Ali Parandeh is a Chartered Engineer (IMechE) and founder of Build Your AI. He has spent 13 years across rail signalling, rail AI and generative AI consultancy, and is the author of Building Generative AI Services with FastAPI (O'Reilly) and the Building Reliable AI Agents liveProject series (Manning).

If getting AI past pilot is a live problem for your team, let's talk.

Sources